Elamur Research⌕
☰
home / papers / 2026 / Lost in Activations: A Neuron-level Analysis of Encoders for Cross-Lingual Emotion Detection

Lost in Activations: A Neuron-level Analysis of Encoders for Cross-Lingual Emotion Detection

Paper · 2026 · Pranaydeep Singh, Orphee De Clercq and Els Lefever

Summary

Multilingual encoders transfer well across languages, but it is unclear whether this rests on shared, language-independent units or on overlapping language-specific ones, and no prior work had looked at emotion-related neurons from both angles. The authors take XLM-R base, fine-tuned only on the English part of the EXALT dataset for six emotion classes, and test it zero-shot on Arabic, Dutch, Japanese and Slovenian. They extract the 768-dimensional [CLS] vector from layer 11, normalize it per language, and run three analyses. First, they rank neurons by activation variance across labels. Second, they compute Spearman correlations of activations between language pairs. Third, they sort neurons into "polyglot" and monolingual groups. Some neurons encode emotions consistently across languages, while others specialize in one language. Cross-lingual alignment tends to follow typological closeness to the other languages and to English. Because the framework needs no retraining or gradient access, it is a cheap way to see where zero-shot transfer works and where it breaks down.

Key points

  • Setup and performance: The model is XLM-R base topped with a single-layer classifier that predicts one of six categories: Joy, Sadness, Anger, Fear, Love, or a Neutral class. Zero-shot macro-F1 is 0.346 (Arabic), 0.433 (Dutch), 0.420 (Japanese) and 0.439 (Slovene).
  • How emotions are encoded: Joy is spread across many neurons with moderate variance. Sadness relies on a few sharply discriminative neurons.
  • Cross-lingual correlations:
    • Dutch–Slovenian is highest (ρ = 0.94).
    • Arabic–Dutch (0.77) and Arabic–Slovenian (0.84) are lower.
    • Japanese–Dutch is unexpectedly high (0.90), possibly because both languages are well resourced in XLM-R's training data.
  • Neuron types: About 15% of emotion-sensitive neurons are polyglot (above a 0.5 threshold in all four languages) and about 35% are monolingual.
    • Monolingual neuron counts are 13 for Slovenian, 98 for Japanese and 111 for Arabic.
    • Polyglot neurons are more common for Fear and Joy and rarer for Love.
  • Limitations and code: The analysis is correlational only, with ablation and causal analysis left as future work. Code is available at github.com/pranaydeeps/emotion_interp.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Ghent UniversityGhent · Belgium25 works · 2,949 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 3Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 worksMultilingual & cross-lingual NLP · 828 works

Cite

Cite as Pranaydeep Singh, Orphee De Clercq and Els Lefever. 2026. Lost in Activations: A Neuron-level Analysis of Encoders for Cross-Lingual Emotion Detection. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-short.9

@inproceedings{singh2026lost,
  title = {{Lost in Activations: A Neuron-level Analysis of Encoders for Cross-Lingual Emotion Detection}},
  author = {Pranaydeep Singh, Orphee {De Clercq} and Els Lefever},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-short.9},
  url = {https://doi.org/10.18653/v1/2026.eacl-short.9}
}

References

21 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Haya Al-Sheikh. 2014. A cognitive study of anger metaphors in arabic and english. Journal of Language Studies, 14(3):57–74. find
  2. Rochelle Choenni and Ekaterina Shutova. 2022. Investigating language relationships in multilingual sentence encoders through the lens of linguistic typology. Computational Linguistics, 48(3):635–672.
  3. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  4. Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Yonatan Belinkov, D. Anthony Bau, and James Glass. 2019. What is one grain of sand in the desert? analyzing individual neurons in deep NLP models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6309–6317. find
  5. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  6. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics, 8:842–866.
  7. Paul Ekman. 1992. An argument for basic emotions. Cognition & Emotion, 6(3-4):169–200.
  8. Ala N. Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025. Mechanistic interpretability of emotion inference in large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 13090–13120, Vienna, Austria. Association for… find
  9. Logan Foote, Neel Nanda, Ján Kramár, Ioannis Konstas, Jonathan Cohen, and Fazl Barez. 2023. Neuron-to-graph: Interpretable representation learning for neurosymbolic reasoning. arXiv preprint arXiv:2305.19911. find
  10. Takeshi Kojima, Itsuki Okimura, Yusuke Iwasawa, Hitomi Yanaka, and Yutaka Matsuo. 2024. On the multilingual ability of decoder-based pre-trained language models: Finding and controlling language-specific neurons. In Proceedings of the 2024 Conference of the North American Chapter of the Association…
  11. Zoltán Kövecses. 2000. Metaphor and Emotion: Language, Culture, and Body in Human Feeling. Cambridge University Press. find
  12. Jaewook Lee, Woojin Lee, Oh-Woog Kwon, and Harksoo Kim. 2025. Do large language models have “emotion neurons”? investigating the existence and role. In Findings of the Association for Computational Linguistics: ACL 2025, pages 15617–15639, Vienna, Austria. Association for Computational Linguistics. find
  13. Weize Liu, Yinlong Xu, Hongxia Xu, Jintai Chen, Xuming Hu, and Jian Wu. 2024. Unraveling Babel: Exploring multilingual activation patterns of LLMs and their applications. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 11855–11881, Miami, Florida, USA… find
  14. Miguel López-Otal, Jorge Gracia, Jordi Bernad, Carlos Bobed, Lucía Pitarch-Ballesteros, and Emma Anglés-Herrero. 2025. Linguistic interpretability of transformer-based language models: a systematic review. Preprint, arXiv:2504.08001. — [arXiv:2504.08001](https://arxiv.org/abs/2504.08001) find
  15. Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, and Ji-Rong Wen. 2024. Language-specific neurons: The key to multilingual capabilities in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu…
  16. Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. 2020. Ccnet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC 2020), pages 4003–…
  17. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Ling…
  18. Die Zhang, Hao Zhang, Huilin Zhou, Xiaoyi Bao, Da Huo, Ruizhao Chen, Xu Cheng, Mengyue Wu, and Quanshi Zhang. 2021. Building interpretable interaction trees for deep NLP models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14328–14337. find
  19. Aaron Maladry, Pranaydeep Singh, and Els Lefever. 2024. Findings of the wassa 2024 exalt shared task on explainability for cross-lingual emotion in tweets. In Proceedings of the 14th Workshop of on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis@ACL 2024, Bangkok, Thailan… find
  20. Saif M. Mohammad and Peter D. Turney. 2013. Crowdsourcing a word–emotion association lexicon. Computational Intelligence, 29(3):436–465. find
  21. Gonçalo Santos Paulo, Alex Troy Mallen, Caden Juang, and Nora Belrose. 2025. Automatically interpreting millions of features in large language models. In International Conference on Machine Learning (ICML). OpenReview/ICML 2025 poster track. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/singh-2026-lost-activations-neuron-level-analysis/